Motivation

Computational methods play a central role in climate research. In the construction of future climate scenarios, for instance, the complexity of the systems involved, the limited feasibility of "climate experiments" and the need to systematically investigate uncertainties in empirical data and/or in mathematical models make computational methods unavoidable.

In spite of the different systems involved in climate research, e.g., ocean and economy systems, climate models are based on a small set of algorithms.

One would expect these algorithms to be implemented in generic libraries that fulfill well-defined specifications and that can be reused and applied through different domains.

This, however, is not the case. Formal problem specification are, in climate research, rare and computational methods are developed compartmentally. Therefore, atmosphere, vegetation and economy models, for instance, are often based on different abstractions, built according to different architectures and implemented in different programming languages. They share virtually no software components.

The lack of shared abstractions, unambiguous problem specifications and generic implementations leads to wasteful usage of intellectual resources, thwarts model coupling efforts and prevents a desirable reduction of model uncertainty through systematic component-based model validation.

The absence of a common base of computational algorithms also hinders climate models to take full advantage of modern parallel computing architectures. Parallel computation, in turn, could allow simulations on longer time scales, improved resolution and more representative sets of realizations in uncertainty studies.

The aim of this project is to conceive, design and implement reusable, generic software components for distributed adaptive finite volume methods. These components should be flexible and efficient enough to support the development of realistic model prototypes and applied research in numerical modelling. They should allow experts to implement their know-how by learning only a few software interfaces and working on well defined abstractions. Practitioners should be enabled to meaningfully combine components into complex models and set up documentable numerical experiments.